When a harness compacts, you lose visibility into what left the context window. It is most noticeable on Codex with the default 250k window, where compaction happens often.
The idea is to mark the compaction point in the transcript, show which messages are no longer in the agent's context, and let you resend anything important that got dropped.
We already keep the full pre-compaction transcript in our own store, so the history is there. What we do not record is the boundary itself, and for Codex auto-compaction we currently record nothing at all.
Showing the actual summary text is only possible on Claude and OpenCode, where it gets produced and we discard it today. Codex encrypts it, and Grok, Pi and omp never emit one. Cursor and fx cannot compact at all. So the boundary view can work almost everywhere, and the summary would be a bonus on the two harnesses that expose it.
When a harness compacts, you lose visibility into what left the context window. It is most noticeable on Codex with the default 250k window, where compaction happens often.
The idea is to mark the compaction point in the transcript, show which messages are no longer in the agent's context, and let you resend anything important that got dropped.
We already keep the full pre-compaction transcript in our own store, so the history is there. What we do not record is the boundary itself, and for Codex auto-compaction we currently record nothing at all.
Showing the actual summary text is only possible on Claude and OpenCode, where it gets produced and we discard it today. Codex encrypts it, and Grok, Pi and omp never emit one. Cursor and fx cannot compact at all. So the boundary view can work almost everywhere, and the summary would be a bonus on the two harnesses that expose it.